# Reuse Normalization Layer in Multiple Models?

**URL:** <https://discuss.ai.google.dev/t/reuse-normalization-layer-in-multiple-models/19208>\
**Category:** Keras\
**Tags:** recommenders, models\
**Created:** [November 17, 2023, 5:46pm UTC](https://discuss.ai.google.dev/t/reuse-normalization-layer-in-multiple-models/19208 "2023-11-17T17:46:40Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![rcauvin](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/rcauvin/32/10360_2.png) [@rcauvin](https://discuss.ai.google.dev/u/rcauvin)\
**Post date:** [November 17, 2023, 5:46pm UTC](https://discuss.ai.google.dev/t/reuse-normalization-layer-in-multiple-models/19208/1 "2023-11-17T17:46:40Z")

</div>

I’m training multiple models from the same input data, and I want each of the models to use the same normalization layer:

> normalizer = tf.keras.layers.Normalization(axis = None)  
> normalizer.adapt(vocabulary)

Can I use the same normalizer instance as a Keras layer in all three models without any transfer learning taking place? Or will training the first model create weights or “state” for the layer that will transfer to the subsequent models that share the same layer?

(The context here probably doesn’t matter, but I’m using TensorFlow recommenders, where I am training two retrieval models and a ranking model.)

---

<div class="post-metadata">

**Author:** ![tagoma](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/tagoma/32/19434_2.png) [@tagoma](https://discuss.ai.google.dev/u/tagoma)\
**Post date:** [November 19, 2023, 9:19am UTC](https://discuss.ai.google.dev/t/reuse-normalization-layer-in-multiple-models/19208/2 "2023-11-19T09:19:23Z")

</div>

I think one can say **Normalization** is a preprocessing layer that has an internal state and is non-trainable\*. The state of **Normalization** layers are set before training .

You provided hardly few elements on you model, but maybe you can add to each of you model a second **Input** layer that would be fed by a bucket of all the data you want to normalize, hence followed by a **Normalization** layer, and the rest of your model would be fed with data from the other **Input** layer? There are probably many ways to get what you want anyways.

\*It can be “adapted” though